Alpha Arena Season 1 (6 LLMs, real money on Hyperliquid)

0 of 6 reach |t| = 22026-10-15Sources public, fetched on the stated date

Finding

No model is distinguishable from luck in either direction: the winner's t is 0.54, the worst loser's −0.83. 17 days needs a Sharpe of about 9.3 to reach t = 2.

The subject is the whole season, not any one model.

Subject

Nof1's Alpha Arena, Season 1: six large language models (Qwen3 Max, DeepSeek Chat V3.1, Claude Sonnet 4.5, Grok 4, Gemini 2.5 Pro, GPT-5), each given $10,000 of real money, traded crypto perpetuals on Hyperliquid from 2025-10-17 to 2025-11-03 with the same prompt and data. Qwen3 Max finished first; GPT-5 lost more than 60%. The result was widely reported as one model being better at trading than another.

Three inputs

Input	Value
Knowledge cutoff	Not applicable. Live trading is closed-book by construction; that is a strength of the format.
Window	17 days (2025-10-17 22:00Z to the close, 2025-11-03 22:00Z).
Attempts	6 (models in the season); the winner is the best of 6.

Finding

None of the six results is statistically distinguishable from luck, in either direction.

Data: hourly dollar_equity per model from Nof1's own api/account-totals, as archived by the Wayback Machine on 2025-11-12 (the live endpoint and the arena pages returned 404 on 2026-10-12). Every model is anchored at its $10,000 start.
Method: daily returns → the estimator behind the wallet check (Sharpe × √years, Student-t on effective degrees of freedom, put back on the normal scale). Script: tools/audit_alpha_arena.py.
Models with |t| ≥ 2: 0 of 6. The winner's t is 0.54; the worst loser's is −0.83.
Picking the best of 6 raises the bar from t = 2 to t ≈ 2.67 (Bonferroni, one-sided).

Model	Return	Sharpe (daily)	t
Qwen3 Max	+20.6%	2.56	0.54
DeepSeek Chat V3.1	+4.0%	1.95	0.41
Claude Sonnet 4.5	−31.9%	−2.02	−0.43
Grok 4	−45.4%	−3.32	−0.70
Gemini 2.5 Pro	−55.2%	−2.79	−0.59
GPT-5	−62.1%	−3.95	−0.83

Returns are from the last archived hourly sample at the close and differ slightly from the figures Nof1 announced (for example +22.3% for Qwen3 Max); the finding does not depend on the difference.

Why the window decides it

t = Sharpe × √years. Over 17 days, √years ≈ 0.22, so:

to reach t = 2, a model needs an annualised Sharpe of about 9.3;
to clear the best-of-6 bar, about 12.4.

Very few trading records sustain either. A 17-day season can rank models; it cannot tell skill from luck, and the losers have not been shown to be bad any more than the winner has been shown to be good.

What it would take

At the season's own Sharpe, if it held:

Model	Sharpe	Months for t = 2	Months for best of 6
Qwen3 Max	2.56	about 7	about 13
DeepSeek Chat V3.1	1.95	about 13	about 23

Worth noting

Nof1's own homepage now says current models are poor at markets and improve little between releases, which is consistent with this finding. Live, closed-book trading is the right format for testing models; what it needs is a season long enough to carry the evidence, and a bar that allows for how many entrants there were.

Recheck: when a new season publishes its equity series.